Help the agent know what is important, what is not —
and prevent it from making silent decisions.

A conversation with Isaac Kargar — Nazmi

Interview by Nicolas Dolenc · 2026-08-04

Isaac Kargar has spent a decade building AI systems that have to survive contact with a customer — and the last year building the thing that watches what those systems decide when nobody is looking. He has a name for the failure mode that worries him, and it isn't a wrong number.

How does your organisation decide what to trust in its product data — and when does a bad number become a decision that costs you?

In one line

For Isaac Kargar, trust isn't a property of a number. It's a property of the harness that decided what to extract, what to keep and what to hand to an agent unsupervised. The dangerous failure isn't a wrong figure — it's an unverified decision nobody noticed was made.

Raw traces aren't data until something filters them.

“There are a lot of data — chat traces, agent traces — you want a context layer on top of this to filter and extract the decisions made, the different signals, from those traces.”

Chat logs, agent traces, Slack and Teams history are noise until a layer extracts the decisions and signals worth keeping. The filter is not a preprocessing step; it is the system.

Trustworthy memory is defined by what it drops.

“The hardest part, especially for memory, was finding the balance — not memorizing too many details, but also not failing to memorize what matters.”

Over-retention is as corrosive as under-retention. His answer is a hierarchy — recent context held close, mid-term facts merged, unused items decayed, the rest parked in a database and retrieved only on demand.

Predictability beat autonomy.

“In the end I came to the conclusion that a simple workflow can be enough — several components, LLM calls, do this, then do this — much more predictable, easier to build, and you can use smaller models.”

He built the fully agentic version, then a self-pruning harness, then discarded both. Decomposition is what makes a system inspectable, and inspectable is what makes it trustworthy.

The expensive failure is the silent decision.

“One of the things I try to do is help the agent know what is important, what is not — to try to prevent it from making silent decisions.”

Not a wrong number on a dashboard — an unverified intermediate choice that compounds until the output looks fine and isn't. His concrete case: “I found about 1,000 data points in the set were missing information the agent needed... So I started breaking the steps down: the agent does the step, but I verify the result before continuing.”

Trust drifts upward on its own, so it has to be reset deliberately.

His own observation about working with models over time: confidence in a model rises with familiarity while measured performance appears to fall — because the checking stopped. Media enthusiasm accelerates the drift. Resetting expectations is maintenance, not pessimism.

Systematising by hand is the price of safe delegation.

“In a startup you often have to do the hard work yourself first, which isn't scalable — you convert it into systematized steps, and only then can an AI agent maybe do a better job.”

You cannot hand over work you have not yet decomposed. The unscalable version is not waste; it is the specification.

What survives is the expertise around the instruction.

“Many people can run [an agent] and tell it to do even the most difficult things... but those expertise details — how to collect the data, clean the data, how to train, what is the role of this component, how to make it better — maybe AI gets there, but I don't know we can do much more [without it].”

He expects companies to get smaller. He does not expect the judgment layer to disappear.

Help the agent know what is important, what is not —and prevent it from making silent decisions.